A comprehensive and hybrid approach to automatic and interactive point cloud segmentation using surface variation analysis and HDBSCAN clustering

Abstract

Segmentation is one of the four main operations involved in processing point clouds for reverse engineering and metrology. Numerous segmentation methods exist, including region growing, attribute clustering, edge detection, and machine learning-based approaches, each with its own strengths and weaknesses. Hybrid approaches, which combine these methods, can often yield improved results. This paper proposes a novel, hybrid, four-step method for segmenting 3D point clouds of mechanical parts, based on surface variation analysis and a clustering technique. The method begins with the evaluation of a surface variation parameter to differentiate edge and non-edge points, followed by threshold-based separation of the edge points. An edge point expansion technique is then introduced to improve segmentation results by enhancing the spatial distinction between edge and surface points, thereby minimizing the sensitivity of the clustering algorithm. Finally, the HDBSCAN clustering method is employed to group the remaining points into distinct clusters representing individual surfaces. The effectiveness of the proposed technique is validated through experiments on synthetic point clouds of mechanical parts, incorporating added noise and density variations. These experiments demonstrate the method's robustness in reverse engineering applications for mechanical components. A measured point cloud is also taken as an example to verify the feasibility of the proposed method. An interactive graphical user interface (GUI) is also developed to facilitate real-time adjustments during the segmentation process. This research significantly contributes to automatic 3D point cloud analysis and supports advancements in Industry 4.0.